Bedrock Claude 3.x Shutdown on November 13: Which Claude Model to Move To

Last updated: October 18, 2026

Amazon Bedrock is ending the last Claude 3.x models in the Regions where they still run. If your code calls anthropic.claude-3-haiku-20240307-v1:0, anthropic.claude-3-sonnet-20240229-v1:0, anthropic.claude-3-5-sonnet-20240620-v1:0, anthropic.claude-3-5-sonnet-20241022-v2:0 or anthropic.claude-3-7-sonnet-20250219-v1:0 in one of seven AWS Regions, those requests stop working on November 13, 2026. That is 26 days from today. This post covers the exact dates and Regions, what happens to requests, which model to move to, what it costs, the new model ids and the code change. Everything below was checked against AWS's and Anthropic's own docs on October 18, 2026.

The dates, by Region#

The schedule is on Bedrock's model lifecycle (Legacy) page, which covers every model launched on Bedrock before September 7, 2026. The newer model lifecycle page only applies to models launched on or after that date, so it does not list these. All five rows end on the same day:

The seven Regions are Seoul (ap-northeast-2), Mumbai (ap-south-1), Singapore (ap-southeast-1), Zurich (eu-central-2), London (eu-west-2), Sao Paulo (sa-east-1) and Canada Central (ca-central-1). For every row, the page puts the "public extended access start date" on the same day as the Legacy date, so these models are already in extended access.

Claude 3 Haiku has already gone once. It reached end of life on September 10, 2026 in us-east-1, us-east-2, us-west-2, eu-central-1, eu-west-1, eu-west-2, eu-west-3, ap-northeast-1, ap-southeast-2 and both GovCloud Regions. The Claude 3 Haiku model card says requests there "now fail with ResourceNotFoundException: This model version has reached the end of its life." That is the error to expect from November 13 in the Regions above.

Two oddities on AWS's side. First, the same lifecycle page says "A model will be in the Legacy state for at least 6 months before the EOL date." These rows give five weeks or less, and the page does not explain the gap. Second, the Claude 3 Haiku model card's warning box is behind the lifecycle page: it lists only Seoul as Legacy, while the lifecycle table also puts Mumbai and Singapore (October 10) and Zurich, Sao Paulo and Canada Central (October 13 and 15) on the November 13 date. AWS settles which one wins: "For Amazon Bedrock usage, only the dates on this page apply." Plan against the lifecycle table.

The table also only lists "the Regions covered by a model's announced EOL date." AWS adds that "the absence of a Region from a row doesn't mean the model will stay available there indefinitely." If you call any Claude 3.x id from a Region not listed here, check its state in that Region, as shown in the checklist below. Our Bedrock deprecations page tracks these rows alongside the rest of the Bedrock schedule.

What Legacy and EOL mean for your requests#

These come from the lifecycle page, quoted where the wording matters:

• Now, in Legacy: "New customers can't use Legacy models and existing customers may lose access to Legacy models after 15 days of inactivity." You also cannot create new Provisioned Throughput or new fine-tuning jobs. Existing on-demand custom model deployments and Provisioned Throughput created before the model entered Legacy can keep running.
• Extended access: active users keep access until EOL, but "you should expect higher pricing, which will be set by the model provider." We did not find a separate extended-access price for these rows on the Bedrock pricing page today, so watch your bill.
• From November 13, at EOL: "requests made to this version will fail, unless there is a private arrangement between you and the provider for continued access." And: "Migration will not happen automatically."

So there is no redirect. If a fallback path in your code uses one of these ids and has not sent traffic in 15 days, it may already be gone.

Which surface each date applies to#

Every date above is a Bedrock date. On Anthropic's own Claude API, all five models are already retired, according to Anthropic's model deprecations page:

Anthropic says its dates "apply to Anthropic-operated platforms: the Claude API, Claude Platform on AWS, and Microsoft Foundry," and that "Partner-operated platforms (Amazon Bedrock and Google Cloud) set their own retirement schedules." So if any of these ids still answers for you, you are almost certainly calling it through Bedrock, and November 13 is your last day there. For the cross-platform calendar, see our AI model shutdown dates for 2026-2027.

Which model to move to#

AWS's lifecycle page names no replacement for any of these models. It says only to "migrate to the latest model." AWS's models at a glance index no longer has cards for Claude 3 Sonnet, 3.5 Sonnet or 3.7 Sonnet. Under Claude 3.x it lists Claude 3 Haiku alone.

The replacements in the table above are Anthropic's, for its own API. They are a reasonable starting point on Bedrock too. The table below is our suggestion, built from the Claude models AWS lists as Active on Bedrock today:

Why these defaults. Sonnet 5.5 is Anthropic's named replacement for every Sonnet on the list, and its Bedrock card gives it an "EOL no sooner than" date of September 28, 2027. Haiku 5.5 launched on Bedrock on October 7, 2026, after Anthropic wrote its Haiku 3 notice, with an EOL floor of October 7, 2027. Anthropic's pick, Haiku 4.5, still works. But its Bedrock card shows an EOL floor of October 16, 2026, which has already passed, so it could enter Legacy at any time with notice. Moving to it may mean migrating twice. Sonnet 4.6, the Sonnet alternative, has an EOL floor of February 17, 2027 on its card.

Why the alternatives exist. The Bedrock model cards list features per model, and two gaps matter for code written against Claude 3:

• Forced tool calls. The Sonnet 5.5 card supports client-side tool calling with "tool_choice types auto and none only." If you set tool_choice to a specific tool or to any, a common way to get JSON out of Claude 3, that will not carry over. The Sonnet 4.6 card lists no such restriction.
• Structured outputs. The Sonnet 5.5 and Haiku 5.5 cards list structured outputs as not supported. The Sonnet 4.6 and Haiku 4.5 cards list them as supported on bedrock-runtime. One exception: Haiku 4.5 does not support them through the India profile (in.anthropic.claude-haiku-4-5-20251001-v1:0).

The step up in limits is large either way. The cards give Claude 3 Haiku a 200K context window, 4K max output and an August 2023 knowledge cutoff. Sonnet 5.5 and Haiku 5.5 have 1M tokens of context, 128K max output and a June 2026 cutoff. Haiku 4.5 has 200K, 64K and February 2025; Sonnet 4.6 has 1M, 64K and August 2025. Our knowledge cutoff and context window tables compare these across providers, and the Anthropic model hub has every Claude id in one place.

Check your Region first: the data residency catch#

This is the part most likely to block a quick swap. On the bedrock-runtime endpoint, the Claude 3 Haiku card shows In-Region support in six of the seven Regions. In-Region means requests stay in the one Region you call. The replacement cards mostly do not offer In-Region in these Regions. Here is what each card shows for the seven Regions on bedrock-runtime today:

What that means in practice:

• Seoul, Singapore and Sao Paulo: none of the four cards offers an In-Region or Geo option. The only path is a Global profile, whose destinations, per AWS's inference profile docs, "include all commercial Regions."
• Zurich: the eu. profile called from eu-central-2 routes to Frankfurt, Zurich, Stockholm, Milan, Spain, Ireland and Paris, so requests can leave Switzerland.
• Canada Central: the us. profile called from ca-central-1 routes to ca-central-1, us-east-1, us-east-2 and us-west-2.
• Mumbai: only Haiku 4.5 has an India geography profile, in.anthropic.claude-haiku-4-5-20251001-v1:0, routing between Mumbai and Hyderabad.
• London: Sonnet 4.6 is the only one of the four with In-Region support.

If a residency requirement is why you call Claude from one of these Regions, take this table to whoever owns that requirement this week, not on November 12. AWS's own advice: "Don't rely on the geographic prefix in an inference profile ID alone to determine where requests can be routed." Call GetInferenceProfile from your source Region and read the destination list. For more on what crossing the AWS boundary means, see Fable 5 on AWS Bedrock: When Your Data Leaves the AWS Boundary.

Price comparison#

All prices are per million tokens, read from the Bedrock pricing page on October 18, 2026. The page showed its default region, US East (Ohio). We could not read the region selector, so confirm the rates for your Region there. The Claude 3.x rows appear only in the "Geo and In-region" table. The newer models appear in both, with Geo and In-region priced 10% above Global:

The Global rows match Anthropic's first-party prices for the same models. So what does each move cost?

• Sonnet 3.x to Sonnet 5.5: cheaper. Input drops from $3.00 to $2.00 Global ($2.20 Geo), output from $15.00 to $10.00 ($11.00). The pricing page also notes that from October 7, 2026, Sonnet 5.5 cache reads were cut 50%.
• Sonnet 3.x to Sonnet 4.6: the same $3.00 / $15.00 on Global, 10% more on Geo or In-Region.
• Claude 3 Haiku to Haiku 5.5: cheaper per token, $0.10 / $0.50 against $0.25 / $1.25. One catch from the Haiku 5.5 card: prompts over 100,000 tokens bill at a separate "Long Context" rate, $0.50 / $2.50 Global. Claude 3 Haiku could not take a prompt that long anyway.
• Claude 3 Haiku to Haiku 4.5: four times the per-token price, $1.00 / $5.00. That is Anthropic's first-party pick, and the main reason we default to Haiku 5.5 instead.

Per-token price is not cost per task. Sonnet 5.5 and Haiku 5.5 have adaptive thinking on by default (per their cards), and thinking tokens bill as output. Measure on your own traffic before you commit to a budget.

Model id changes#

The old ids were base model ids called in-Region. The new ones are mostly inference profile ids, with a prefix that sets where requests can go. From each AWS model card:

Three things to notice:

• The 5.5 and Sonnet 4.6 ids drop the date stamp and the -v1:0 suffix. If you parse model ids with a regex that expects -YYYYMMDD-v1:0, fix the regex.
• For Haiku 4.5, the card is explicit: "the bedrock-runtime endpoint requires a geo or global inference profile ID. The bare model ID isn't supported for on-demand throughput." The Sonnet 5.5, Haiku 5.5 and Sonnet 4.6 cards show no In-Region support in the affected Regions except Sonnet 4.6 in London, so use a profile id there too.
• Cross-Region profiles need IAM and SCP permissions in every destination Region. AWS warns: "If any destination Region in a cross-Region inference profile is blocked in your SCPs, the request will fail even if other Regions remain allowed." Organizations that lock Bedrock to one Region will hit this on the first request.

The code change#

Here is the Converse example from AWS's Claude 3 Haiku model card:

import boto3

client = boto3.client('bedrock-runtime', region_name='us-east-1')
response = client.converse(
    modelId='anthropic.claude-3-haiku-20240307-v1:0',
    messages=[
        {
            'role': 'user',
            'content': [{'text': 'Can you explain the features of Amazon Bedrock?'}]
        }
    ]
)
print(response)

The swap to Haiku 5.5 is one line. This is the id AWS's Haiku 5.5 card uses in its own Converse sample:

 response = client.converse(
-    modelId='anthropic.claude-3-haiku-20240307-v1:0',
+    modelId='global.anthropic.claude-haiku-5-5',
     messages=[

Keep region_name set to the Region you call today. That Region becomes the source Region for the profile, and the cards list Global support from all seven. For a Sonnet, the id is global.anthropic.claude-sonnet-5-5, or a geo id from the table above. On InvokeModel, the request body does not change. This is AWS's Sonnet 5.5 sample, which keeps the same anthropic_version value Claude 3 used:

import json
import boto3

client = boto3.client('bedrock-runtime', region_name='us-east-1')
response = client.invoke_model(
    modelId='global.anthropic.claude-sonnet-5-5',
    body=json.dumps({
            'anthropic_version': 'bedrock-2023-05-31',
            'messages': [{ 'role': 'user', 'content': 'Can you explain the features of Amazon Bedrock?'}],
            'max_tokens': 1024
    })
)
print(json.loads(response['body'].read()))

Both snippets are copied from AWS's model cards with only the model id changed, and both pass a Python syntax check. We made no live Bedrock calls while writing this, so run them against your own account.

Before you ship the swap, check four request settings:

• Sampling parameters. Anthropic's deprecations page says non-default temperature, top_p or top_k "Returns a 400 error when set to a non-default value on Claude 4.7 and later models." Claude 3 code often sets temperature, including through Converse's inferenceConfig. The page does not say whether Bedrock behaves the same, so test with and without it.
• Thinking. Sonnet 5.5 defaults to adaptive thinking at effort high. Haiku 5.5 defaults to it at medium, and its card says thinking "can be turned off," with effort capped at high when it is. Expect more output tokens and latency than Claude 3 gave you unless you tune this. The Sonnet 5.5 guide and Haiku 5.5 guide cover the effort settings.
• tool_choice and structured outputs. These are the gaps covered above. Grep for tool_choice and toolChoice before you pick a model.
• Prompt caching. The minimum tokens per cache checkpoint is 512 on Sonnet 5.5 and Haiku 5.5, 1,024 on Sonnet 4.6 and 4,096 on Haiku 4.5. A cache breakpoint that worked on a short Claude 3.x prompt may silently not cache on Haiku 4.5.

Checklist#

• Grep your code, config and infrastructure-as-code for claude-3-haiku-20240307, claude-3-sonnet-20240229, claude-3-5-sonnet-20240620, claude-3-5-sonnet-20241022 and claude-3-7-sonnet-20250219. Include prefixed forms such as us.anthropic. and apac.anthropic..
• For each Region you call, run ListFoundationModels or GetFoundationModel and read the modelLifecycle field. AWS names this as the way to check the current state per Region.
• Send traffic on any fallback path that uses these ids, or remove it. Legacy access can lapse after 15 days of inactivity.
• Pick a target per call site: Sonnet 5.5 or Haiku 5.5 by default, Sonnet 4.6 or Haiku 4.5 if you rely on forced tool_choice or structured outputs.
• Check the Region table above against your residency requirements before choosing a global. or geo profile.
• Call GetInferenceProfile from your source Region and confirm the destination Regions. Open IAM and SCPs for all of them.
• Remove non-default temperature, top_p and top_k, or test that they are accepted.
• Decide on thinking and effort settings, then re-run your evals on the new model.
• Confirm prices for your Region on the pricing page, and set a budget alert for the first week.
• Ship before November 13, 2026. After that date, requests fail and nothing switches over for you.

Continue Reading#

• AI Model Shutdown Dates 2026-2027: Calendar and Migration Plan - every dated shutdown across OpenAI, Anthropic, Google, Bedrock and Foundry, with a cross-platform checklist.
• Claude Sonnet 5.5 Developer Guide: Pricing and API Changes - the parameter and thinking changes behind the Sonnet swap.
• Claude Haiku 5.5: Pricing, Migration Changes, and When to Use It - the Haiku side of the move, including effort controls.
• Fable 5 on AWS Bedrock: When Your Data Leaves the AWS Boundary - read this before you switch to a Global profile.
• Azure OpenAI and Foundry Retirements in November 2026: o1, o1-pro, codex-mini, Claude Haiku 4.5 - Claude Haiku 4.5 leaves Microsoft Foundry two days later, on November 15.
• MiniMax M2.5 and M2.7 Retirement: How to Move to M3 Before November 18 - another November deadline, five days later.

Sources#

All accessed October 18, 2026.

• Amazon Bedrock model lifecycle (Legacy): dates, Regions, Legacy and EOL rules
• Amazon Bedrock model lifecycle (models launched on or after September 7, 2026)
• Amazon Bedrock models at a glance
• Bedrock model card: Claude 3 Haiku
• Bedrock model card: Claude Sonnet 5.5
• Bedrock model card: Claude Haiku 5.5
• Bedrock model card: Claude Haiku 4.5
• Bedrock model card: Claude Sonnet 4.6
• Supported Regions and models for inference profiles
• Amazon Bedrock pricing
• Anthropic model deprecations
• Claude on Amazon Bedrock (Opus 4.6 and earlier), Anthropic docs